SBIR-STTR Award

Reliable automated pattern recognition in the current system is limited by excessive noise and clutter combined with insufficient and/or ambiguous features. Our team consisting of AcornSI and Leidos thus propose a multi-step approach based on the combined effects of improved detection, feature extraction, phase identification, and global association.A key innovation here is creating a new classification feature and confidence based on machine vision to directly classify waveform spectrograms.Improved legacy and new features are then sent to a (possibly nested) support vector machine (SVM) to automatically determine event classifications with associated confidence scores.machine learning,classification,deep convolutional neural network (DCNN),nuclear explosion,Support Vector Machine (SVM),feature extraction,seismic hydroacoustic infrasound (SHI)

We propose a multi-step approach to event classification using machine learning based on convolutional neural networks for noise reduction, improved phase picking and combined new and legacy features.